METHODS AND APPARATUS FOR DEEP LEARNING NETWORK EXECUTION PIPELINE ON MULTI-PROCESSOR PLATFORM

• US 20190324759A1
• Filed: 04/07/2017
• Published: 10/24/2019
• Est. Priority Date: 04/07/2017
• Status: Active Grant

• Current Assignee: Intel Corporation
• Original Assignee: Intel Corporation
• Inventors: Yang, Liu, Yao, Anbang
• Granted Patent: US 11,461,105 B2
• Time in Patent Office: Days
• CPC Class Codes: G06F 15/80 comprising an array of processing units with common control, e.g. single instruction multiple data processors G06F15/82 takes precedence ; for correlation function computation G06F17/15
• CPC Class Codes: G06F 9/3867 using instruction pipelines
• CPC Class Codes: G06F 9/3893 controlled in tandem, e.g. multiplier-accumulator
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 3/044 Recurrent networks, e.g. Hopfield networks
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/063 using electronic means
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06N 3/084 Backpropagation, e.g. using gradient descent
• CPC Class Codes: G06N 3/088 Non-supervised learning, e.g. competitive learning
• CPC Class Codes: G06T 1/20 Processor architectures; Processor configuration, e.g. pipelining
• CPC Class Codes: G06T 15/005 General purpose rendering architectures

Abstract

Methods and systems are disclosed using an execution pipeline on a multi-processor platform for deep learning network execution. In one example, a network workload analyzer receives a workload, analyzes a computation distribution of the workload, and groups the network nodes into groups. A network executor assigns each group to a processing core of the multi-core platform so that the respective processing core handle computation tasks of the received workload for the respective group.

Claims

1. A pipeline framework to run a deep learning network having a plurality of network nodes on a multi-core platform, the pipeline framework comprising:

a network workload analyzer to receive a workload, to analyze a computation distribution of the workload, and to group the network nodes into groups; and

a network executor to assign each group to a processing core of the multi-core platform so that the respective processing core handles computation tasks of the received workload for the respective group.

2. The framework of claim 1, wherein the network workload analyzer groups the network into a number of groups equal to a number of cores of the multi-core platform.

3. The framework of claim 1 or 2, wherein the network workload analyzer groups the network nodes based on an amount of computation for each group.

4. The framework of claim 3, wherein the network workload analyzer distributes a similar amount of workload to each group.

5. The framework of any one or more of the above claims, wherein the deep learning network is configured as an acyclic graph and the flow of data through each network node flows in a single direction through the graph and wherein the data flow through each group is in the same direction for each group.

6. The framework of any one or more of the above claims, wherein the workload comprises a sequence of consecutive images and wherein each image is assigned to multiple groups.

7. A multi-processor computation method comprising:

receiving a workload for pipeline framework to run on a deep learning network having a plurality of network nodes;

analyzing the computation distribution of the workload across the network nodes;

grouping the network nodes into groups based on the computation distribution;

assigning each group to a respective processing resource of the pipeline framework; and

executing the workload on the processing resources based on the assignments.

8. The method of claim 7, further comprising providing an output of the execution to further processes.

9. The method of claim 7 or 8, wherein the processing resources comprise cores of a multiple core processor.

10. The method of claim 9, wherein each respective processing core handles the computation tasks of the received workload for the respective group of network nodes.

11. The method of claim 7, wherein grouping the network nodes comprises grouping the network nodes in order to distribute the workload evenly across the groups.

12. A deep learning system comprising:

a sensor array to capture a sequence of images;

an input buffer to receive the sequence of images;

a network executor having a plurality of cores to execute a deep learning network on the plurality of cores; and

a network workload analyzer to receive the sequence of images as a workload, to analyze a computation distribution of the workload, and to group the network nodes into groups, wherein the network executor assigns each group to a respective core and executes the groups using the respective core.

13. The deep learning system of claim 12, wherein the network executor is further to allocate a shared memory as an output to one of the plurality of cores and an input to another of the respective cores for sequential execution of each respective group of network nodes.

14. The deep learning system of claim 12 or 13, wherein the network workload analyzer groups the network nodes based on an amount of computation for each group.

15. The deep learning system of any one or more of claims 12, wherein the deep learning network is configured as an acyclic graph and the flow of data through each network node flows in a single direction through the graph and wherein the data flow through each group is in the same direction for each group.